[archontes]

It's a stretch to call training an AI creating a 'derivative work' by the legal definition.

You could count the words in a book and publish the word count, and while the information is based on the contents of the book, that would fall incredibly short of being a derivative work.

I suspect they committed whatever copyright violation is committed when they downloaded the copyrighted works. Training an AI on them is simply not related to the protections that copyright offers.